1%客户支撑的AI泡沫：Token贬值与万亿美元债务账单

1%. We've got one of the biggest build-outs in American history [music] counting on 1% of customers who may switch to a cheaper model, probably Chinese, tomorrow. We made two viral videos about the economics of the AI industry, the circular economy that moves money back and forth between the same players, and how pension funds and life insurance participate in development of data centers. Today is the second half of that story, the market of buyers, those of us who pay for AI. By the end of the video, you will understand how to read the state of the AI market today and see it for what it is. To do that, I'll show you how to follow the money, the tokens, the buyers, and the debts.

Let's dive in.

Also, there is no such thing as a universally correct number of tokens for a task. If you look at the benchmarks, artificial analysis gave every model identical tasks. And GPT-4 Astra used 60 million tokens. Grok 4.7 used 240. Opus 5.5, 260 [music].

And you can't predict this before you run a prompt. Therefore, how many tokens you get doesn't really matter. The single number that really matters is what it takes to run a successful task. Counting all the failed attempts, re-prompting, new chats that you have to spin up, agent steps, tools, review, all of that combined to achieve a goal. That's what matters. So, you're going to spend however many tokens it takes to complete a task, and if it's a ton, so be it. But, the cost of tokens is collapsing. It's almost like when you hire a teenager to shovel snow from your driveway, and the teenager says, "I'll do it for $50 an hour." You look at your driveway and go, "Okay, I think this is about 15 minutes of work. So, yeah, I'll take it." And you find yourself thinking, "Why didn't I just do it myself?" So, when the cost of the unit, of the token [music], that is the measuring stick of the whole industry, is going down, the same money buys more.

Well, on one hand, that's great. But, what about those trillions of dollars spent on the AI buildout? What about those debts and loans and special purpose vehicles? If the cost is collapsing, what's going to happen to the financial modeling behind the AI industry that all the assumptions were built upon?

All the predictions of profits were based on the financial modeling that was partially based on the price of tokens, but the cost of tokens is going down.

Which brings me to my first point.

I haven't clicked through any menu in any app in about 6 months, and I love it. Somewhere in the last year, every app I use grew an AI chat. My email, my notion, my analytics, everything I do at work, I've completely stopped clicking through menus. I open apps and I find the quickest possible way to get to the point. So, I just tell it what to do. Find this, link it to that, send it to them, respond in the morning. And you know what's even worse? Laziness breeds laziness. When I close the apps and look at my personal Mac, it just stares back at me.

My computer, the thing all those apps live on, still makes me do everything by hand. I don't want to be looking for a file cuz I don't remember what I named it. I don't want to look for files to delete cuz my storage is full. Same with my phone.

That is why MacPaw built an A. An A is an autonomous AI assistant that lives on the Mac and the iPhone [music] and works across all of your apps. It can find last week's receipt for my dental appointment, rename it, save [music] it to my drive, and submit it to my insurance company all in one go. It only touches what I give it permission to touch. So, I can still check before anything gets sent. An A is free for 14 days with 10,000 AI credits on M-series Macs and iPhone. The link is in the description. Last but not least, it was built by the team with exceptional taste. [music] It is simple, sleek, incredibly smooth, and a pure joy to use. [music] Please keep making beautiful tech, the MacPaw team, so I can be lazy forever.

This is how many tokens were used back in August, and this is how much money was spent. Now look here. This is September, and the number of tokens going up significantly.

But the price paid for those tokens remains flat.

This is what token inflation means. If you could buy 1,000 tokens hypothetically with $10 in August, now the same $10 buys you 10,000 tokens. But models are getting better. Labs are spending enormous amount of money on building new, but the spending is not paying off. You know what this looks like? Your phone plan. 10 years ago, your plan gave you a couple of gigs of data. Now for the same price, you get unlimited data. Carriers spent billions building 5G. The data got faster, better, more reliable, it's cheaper, but did you start paying more?

No.

Your bill is about the same. That's the devaluation I'm talking about. This is the problem because the entire projection for the AI industry was based on the financial modeling that was based on the assumption that the token price stays the same. But once you see this data with your own eyes, you start noticing something about the way labs talk. Google recently said that they went from about 10 trillion tokens a month to over 3 quadrillion in 2 years. 330 times difference.

But notice what it is that they're bragging about. The number of tokens, not revenue per token. Because when the price of the unit is collapsing, [music] the only thing you can brag about is the volume of that unit. Because it looks great on stage. [music] And here's another thing.

Who dictates the price of a token?

There is a lot of margin baked into this price.

You know whose margin?

Nvidia's.

How do you I know that?

Deep Seek's competitors offer discounts of 30 to 40%. If they can offer a discount of 40% and still make profit, that gives you an estimate of what the margin is. Cloud Code and Codex resellers sell tokens at discounts of about 90%. When your product can be resold at 90% off, the price on the label clearly isn't what it cost to make it.

So, to sum it up, the cost of tokens is going down. The margin is going down, which again, on one hand sounds like a great deal cuz for $20 last year, you could build a landing page. Now for $20, you can get a full-blown app. Yes, I am exaggerating. But when you're a frontier lab, you have to offset collapsing margin. So, how do you do that? You have to grow volume and get more people, more businesses to use your model, which brings me to the opposite side of the equation [music]. The buyer, the one who buys access to the model. We talk about the demand for AI as if it's endless. But, is it? Is there a really a blue ocean of all those people and businesses who have just haven't bought the model yet? Or perhaps, there is not that many people left to sell to.

And finally, red flag number three, models have no loyalty. It's becoming more and more obvious that businesses have no problem switching between models. If there is a Chinese open-source model that gets the job done, and let us not forget that Chinese models were distilled from American versions. Consumers and businesses will switch in the blinking of an eye and won't notice that much difference. Even American government website is using Qwen in their chatbot.

Let alone a regular business owner.

Model loyalty is like cars and gas stations. Do you care whose gas is in your tank?

No. You will pull into whatever station is cheapest. That's where AI is heading. When the models are good enough, the only thing left to compete on is the price. This is earth-shaking if you think about it because for the past 4 years, we have been observing the model race. Models were the key to the market. But we have reached a point when for the vast majority of cases, businesses don't need the best frontier models. [music] The current state is enough for the overwhelming majority of business tasks. [music] The strongest use case for AI models is code. This is obvious from adoption among engineering, five coding, and the significant stagnation of the software talent market. And yes, coding is a gigantic slice with plenty of room for growth, but if you take coding out of the equation, we have reached a point where the speed of progress [music] doesn't convert to profits. To sum it up, the entire industry's bet and financial modeling was based on the fact that the better models will be used more and therefore cost more.

But that's not happening, which brings me to my last point.

There's a new paper called financing the AI buildout written for the Brookings Papers on Economic Activity, a very respected academic journals in macroeconomics. It asks a simple question. How much money do AI data centers need to earn to break even? Step one is what it costs to build them. Step two is that the money isn't free. It's like a mortgage. You don't just pay back the price of the house. You pay the price of the house plus the interest. So for each data center that gets built, it has to earn enough to cover the costs plus interest. Put those two together and you get the bill. And the bill is basically this. To break even, data centers built today need to generate about 9.6 trillion dollars in revenue by 2032.

Which is 5 years left. Is 9.6 trillion a lot? It is enormous. To understand how enormous, let me put this in perspective. That revenue would be 9% of the entire US economy. 9% is roughly the size of all American manufacturing combined. For comparison, the entire finance and insurance industries is 8.1. Real estate and rentals are 13.7. So AI would need to become one of the biggest industries in America in 5 years. Now, does this mean that enormous investments in the infrastructure will never pay off.

No, it doesn't. Just like you can take a 10-year mortgage, and if you can't pay it off, you will extend it to 25 years. Yes, it will take you longer to pay it off, but it doesn't mean you're going to lose the house all of a sudden. Same with AI. [music] What we're going through is a period of severe turbulence. The financials are highly questionable. The value of tokens is going down. Businesses start to be rational about AI spendings. And this is what the real state of the AI market looks like today. It doesn't mean it's all going to It may eventually balance itself out. Plenty of frontier technologies went through the exact same turbulence before. At the same time, let us not forget that we've got a massive industry built on debts, debts and promises that don't seem to add up 4 years in.

So, let's go back to the question, who is going to pay the $10 trillion debt? Right now, the honest answer is investors. The typical company spends $16 a month per employee. A tiny group of whales pay almost for everything, and even they can switch to a cheaper model in a day. Businesses pay for better models only when those models justify the investments. And as of today, that's mostly coding.

So, good scenario, the debts will pay off in the next 6 years. Bad scenario, the debts will spread over [music] the years. Worst case, we get over it. I don't know how this whole market situation will balance itself out, but what I do know is that the competition is good.

Let's see how it unfolds. We hope this was helpful, and we'll see you in the next episode.

Bye.